ComfyUI Node
⚗ Alchemica Prompt (idea / list / Ollama + NAG + latent)
Prompt workstation: idea / your prompt / list / string / examples, written by Krea 2's text encoder or Ollama, encoded, with NAG and the start latent. Same socket order as the LoRA stack on the left and as the sampler on the right, with pass-through connectors.
⚗ Alchemica Prompt (idea / list / Ollama + NAG + latent)
- model
- clip
- vae
- latent
- image
- pass_1
- pass_2
- pass_3
- pass_4
- pass_5
- pass_6
- pass_7
- pass_8
- model
- positive
- latent
- negative
- vae
- clip
- prompt
- pass_1
- pass_2
- pass_3
- pass_4
- pass_5
- pass_6
- pass_7
- pass_8
◄modegenerate from idea►
◄text►
◄detail►
◄negative►
◄text_asidea → generate►
◄line0►
◄example▾►
◄seed0►
◄stylerealistic photo►
◄use_ollamafalse►
◄ollama_model(no Ollama model found: start Ollama, then press R)►
◄ollama_urlhttp://127.0.0.1:11434►
◄unload_ollamatrue►
◄show_llm_optionsfalse►
◄words80►
◄creativity0.00►
◄skinsmooth (no body hair)►
◄instructions►
◄max_tokens400►
◄attention_guidancetrue►
◄nag_scale3.0►
◄nag_tau2.5►
◄nag_alpha0.25►
◄nag_blocksall►
◄nag_sigma_start-1.00►
◄nag_sigma_end0.00►
◄structureoff►
◄structure_scale0.00►
◄structure_blocks8-19►
◄seg_blur10.0►
◄structure_sigma_start-1.00►
◄structure_sigma_end0.00►
◄show_resolutionfalse►
◄aspect2:3►
◄megapixels1.50►
◄batch_size1►
◄text_in—►
Category⚗ AlchemicaMente/conditioning
Inputs (50)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | From the LoRA stack (or the loader). Leaves with NAG / structure applied | |
| clip | CLIP | Krea 2 text encoder: always encodes the prompt, and writes it when Ollama is off | |
| mode | COMBO | generate from idea | generate from idea: the LLM writes scene / detail / negative from a few words (any language). improve prompt: the LLM rewrites YOUR prompt, keeping everything in it. as written: your text is used as it is (no LLM). from list: one prompt per line, picked by 'line' (set its control to increment to walk the list). from string: the text arriving on the 'text_in' socket (from another node). example: a built-in example, to start or to test. With list / string / example, 'text_as' says what to do with the picked text. |
| text | STRING | Idea / prompt / list (one per line, or entries separated by a '---' line). {a|b|c} picks one option, __name__ picks a line of wildcards/name.txt | |
| detail | STRING | As written only: rendering of surfaces, skin, fabric, optics. Empty = scene only | |
| negative | STRING | As written: the negative for the NAG. With the LLM: extra items added to the one it writes (e.g. 'glasses, tattoos') | |
| text_as | COMBO | idea → generate | What to do with the text from list / string / example |
| line | INT | 00–99999 | From list: entry number (0 = first; wraps around). 'increment' = next entry at every run |
| example | COMBO | Example mode: the idea used | |
| seed | INT | 00–18446744073709550000 | Picks the wildcard options, and the LLM variation when creativity > 0 |
| style | COMBO | realistic photo | 8 options: realistic photo, portrait, still life / lab, landscape, architecture, cinematic, +2 |
| use_ollama | BOOLEAN | false | Who writes the prompt. Off = Krea 2's own text encoder (no server, sees images). On = a model served by Ollama (needs 'ollama serve') |
| ollama_model | COMBO | (no Ollama model found: start Ollama, then press R) | Installed Ollama models (press R to refresh). For images use a vision model (qwen2.5vl, llava, gemma3...) |
| ollama_url | STRING | http://127.0.0.1:11434 | — |
| unload_ollama | BOOLEAN | true | Frees the Ollama model from VRAM right after writing, so Krea 2 keeps its memory (recommended on 16 GB) |
| show_llm_options | BOOLEAN | false | Shows length, creativity, skin, instructions. Hidden values still apply |
| words | INT | 8030–200 | Length of the SCENE (the detail is 30-50% of it) |
| creativity | FLOAT | 0.000–1.5 | 0 = same text every time. 0.6-0.9 = variations (change the seed) |
| skin | COMBO | smooth (no body hair) | smooth = no body hair (removed from the prompt, added to the NAG negative) |
| instructions | STRING | Extra rules in any language; they win over the built-in ones | |
| max_tokens | INT | 400100–2000 | Upper limit of the answer |
| attention_guidance | BOOLEAN | true | NAG: the negative works at cfg 1. Off = model passes unchanged |
| nag_scale | FLOAT | 3.00–20 | Strength of the negative at cfg 1. Calibrated: 3 |
| nag_tau | FLOAT | 2.51–10 | How far it may push away from the positive |
| nag_alpha | FLOAT | 0.250–1 | Blend with the original attention |
| nag_blocks | STRING | all | DiT blocks (0-27): 'all', '0-13', '5,8,10-12' |
| nag_sigma_start | FLOAT | -1.00-1–1 | -1 = from the profile hand-off (~0.87): removes the negative WITHOUT changing the composition. 1 = from the start |
| nag_sigma_end | FLOAT | 0.000–1 | — |
| structure | COMBO | off | PAG = sharper and more contrast. SEG = blurred queries (halos above 0.3) |
| structure_scale | FLOAT | 0.00-5–10 | — |
| structure_blocks | STRING | 8-19 | — |
| seg_blur | FLOAT | 10.00.5–1000 | — |
| structure_sigma_start | FLOAT | -1.00-1–1 | -1 = from the profile |
| structure_sigma_end | FLOAT | 0.000–1 | — |
| show_resolution | BOOLEAN | false | Shows the size of the start latent. A latent on the 'latent' socket wins |
| aspect | COMBO | 2:3 | 10 options: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, +4 |
| megapixels | FLOAT | 1.500.25–8 | FINAL resolution. 1.5 MP fast, 3-4 MP with two stages |
| batch_size | INT | 11–16 | — |
| vaeopt | VAE | Passes through to the sampler (preview and final image in the node) | |
| latentopt | LATENT | Optional: e.g. a VAE-encoded photo for img2img. Empty = built here | |
| text_inopt | STRING | Text from another node (mode: from string) | |
| imageopt | IMAGE | Visual reference for the LLM (Krea 2's encoder, or an Ollama vision model) | |
| pass_1opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_2opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_3opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_4opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_5opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_6opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_7opt | * | Pass-through: anything that enters here leaves on the same pass | |
| pass_8opt | * | Pass-through: anything that enters here leaves on the same pass |
Outputs (15)
| Name | Type | Description |
|---|---|---|
| model | MODEL | NAG / structure applied: to the sampler |
| positive | CONDITIONING | Scene + detail: to the sampler's positive |
| latent | LATENT | Start latent: to the sampler's latent_image |
| negative | CONDITIONING | To the sampler's negative (used only at cfg > 1) |
| vae | VAE | The VAE that came in |
| clip | CLIP | The text encoder that came in |
| prompt | STRING | Final positive text (for Save Image, captions...) |
| pass_1 | * | Same as the matching pass input |
| pass_2 | * | Same as the matching pass input |
| pass_3 | * | Same as the matching pass input |
| pass_4 | * | Same as the matching pass input |
| pass_5 | * | Same as the matching pass input |
| pass_6 | * | Same as the matching pass input |
| pass_7 | * | Same as the matching pass input |
| pass_8 | * | Same as the matching pass input |